Hadron Identification in Highly Granular Calorimeters

This thesis investigates the potential of highly granular calorimeters to enhance charged hadron identification through novel data representations combined with machine learning techniques. Using an idealized simulation of an extremely granular calorimeter, we establish benchmarks for the maximum achievable performance of such detectors, independent of practical implementation constraints. The analysis follows a two-step pipeline: first, constructing shower representations that capture the essential physical information from calorimeter data, and second, applying suitable machine learning architectures for particle classification. Two distinct representations are studied. A first example is a set of High-Level Features, with classification performed using Boosted Decision Trees. Alternatively, showers are modeled as point clouds, where calorimeter cells are treated as points, and classification is carried out with Deep Sets. These studies are conducted separately, allowing us to assess how the choice of representation influences both the architecture design and the resulting performance. The results provide a systematic evaluation of particle identification accuracy for pions, kaons, and protons, analyzed as a function of detector granularity and particle energy. This benchmark study offers new insight into the information content of highly segmented calorimeters and defines reference points for the achievable performance in charged hadron identification.